Naughty AIs Are Spilling Their Users' Super Personal Chats Onto the Open Web

Is yours in here? Steady Drip A new report from the security firm UpGuard reveals that some NSFW chatbot sites are oozing explicit user chat contents into the open web — and what those chats contain can be disturbing. According to Wired, UpGuard's investigation focused on 400 "exposed" AI services, all of which were built on an open-source AI protocol called llama.cpp. Researchers from UpGuard were able to determine that 117 IP addresses connected to these poorly built services were leaking user prompts into the digital wild, and that three such systems were leaking data from extensive, sexually explicit interactions with erotic chatbots. Over 24 hours, UpGuard collected nearly 1,000 leaked user chats. Disturbingly, five of them centered on child sexual abuse scenarios. Though UpGuard couldn't determine which specific AI sites or services were leaking the exposed prompts, according to Wired, they were able to make out that the sites are host to roleplay-oriented chatbots designed ... Read full article.